Make normalized tensor shortcuts explicit

[connortsui20]

Rationale for this change

Normalized is a physical encoding whose direction is only approximately unit length, so trusting it belongs to scalar-function policy instead of a logical UnitVector dtype.

• Supersedes: Add UnitVector tensor refinement #9421

What changes are included in this PR?

Adds NormMode::{Exact, AssumeNormalized} to cosine similarity and L2 norm. Exact mode measures physical direction norms, while assumed mode retains the current shortcut; both modes persist, and legacy metadata keeps its existing behavior.

What APIs are changed? Are there any user-facing changes?

CosineSimilarity::try_new, L2Norm::try_new, and build_similarity_search_tree now require an explicit NormMode. This adds no logical dtype or Arrow extension type.

[codspeed-hq]

Merging this PR will degrade performance by 7.36%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 2 improved benchmarks
❌ 3 regressed benchmarks
✅ 1951 untouched benchmarks
🆕 12 new benchmarks
⏩ 442 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	column_x_extension_constant[256]	650.4 µs	834 µs	-22.01%
❌	Simulation	column_x_extension_constant[2]	290.9 µs	355.4 µs	-18.16%
❌	Simulation	column_x_extension_constant[32]	230 µs	273 µs	-15.72%
⚡	WallTime	words_gather_scalar[65536]	9.4 µs	8.3 µs	+13.39%
⚡	Simulation	compress_fsst[(500, 64, 8)]	585.9 µs	523.7 µs	+11.88%
🆕	Simulation	normalized_assume[2]	N/A	159.3 µs	N/A
🆕	Simulation	normalized_assume[256]	N/A	84.2 µs	N/A
🆕	Simulation	normalized_assume[32]	N/A	88.2 µs	N/A
🆕	Simulation	normalized_exact[2]	N/A	261.7 µs	N/A
🆕	Simulation	normalized_exact[256]	N/A	113.7 µs	N/A
🆕	Simulation	normalized_exact[32]	N/A	120.4 µs	N/A
🆕	Simulation	normalized_assume[2]	N/A	36.6 µs	N/A
🆕	Simulation	normalized_assume[256]	N/A	25 µs	N/A
🆕	Simulation	normalized_assume[32]	N/A	25 µs	N/A
🆕	Simulation	normalized_exact[2]	N/A	847.6 µs	N/A
🆕	Simulation	normalized_exact[256]	N/A	232.8 µs	N/A
🆕	Simulation	normalized_exact[32]	N/A	267.8 µs	N/A

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ct/norm-compute-mode (76656f1) with develop (9876921)}

Footnotes

1. 442 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩